Medical Image Processing Reliability Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In the medical field, the scarcity and variability of training data for machine learning applications pose challenges due to limited patient numbers and inconsistent data quality created by multiple doctors across different medical institutions, making it difficult to achieve reliable medical image processing.
Innovation Solution
A medical image processing apparatus that includes an acquirer, a reliability setter, and a learner. The acquirer collects training data, while the reliability setter assigns reliability information based on creation situations and creator information, which the learner uses to generate a learned model through weighted training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If training data is collected from multiple medical institutions and doctors, then the quantity of training data increases, but the quality consistency deteriorates
Solution Approach 1:
The patent applies local quality by assigning different reliability weights to different training data items based on their creation context. Each training data item receives a specific reliability score reflecting its quality, allowing the system to treat different data items differently rather than uniformly. This resolves the contradiction by enabling quantity expansion while maintaining quality control through differential weighting.
Solution Approach 2:
The patent changes the parameter of reliability from a binary or uniform attribute to a continuous weighted value. By introducing reliability weights as a variable parameter that can be adjusted based on creation situation and creator information, the system can optimize the contribution of each training data item. This parameter transformation allows the system to accommodate diverse data sources while maintaining overall quality standards.
2Reliability
If training data quality is strictly controlled, then the reliability of learned models improves, but the quantity of available training data decreases
Solution Approach 1:
The patent applies partial action by selectively emphasizing high-quality training data through higher reliability weights while still incorporating lower-quality data with reduced weights. Rather than discarding any data below a strict quality threshold, the system partially utilizes all available data with varying degrees of confidence. This approach maintains model reliability while maximizing the use of available training resources.
Solution Approach 2:
The patent creates a composite training dataset where high-quality and lower-quality data items are combined with different weights. Similar to composite materials in engineering, the final learned model is a weighted composition of multiple data sources with varying quality levels. This composite approach allows the system to leverage the strengths of high-quality data while still benefiting from the quantity provided by lower-quality data.
3Measurement precision
If reliability information is assigned to each training data item, then the quality assessment becomes more precise, but the system complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-defining reliability weight values for different creation situations and creator profiles before actual training data generation. Rather than calculating complex reliability metrics in real-time during model training, the system establishes reliability weights in advance based on metadata such as institution type, doctor experience level, and creation context. This preliminary assignment simplifies the overall system while maintaining precise quality assessment.
Data Source
AI summary
According to one embodiment, a medical image processing apparatus of an embodiment includes an acquirer, a reliability setter and a learner. The acquirer acquires training data created by a creator on the basis of a medical image. The reliability setter sets, to the training data acquired by the acquirer, reliability information based on a creation situation of the training data or information about the creator who created the training data. The learner generates a learned model using the training data according to weighting based on the reliability information set by the reliability setter.